A neural network pre-selection method for lightning very high frequency radiation source positioning signals
Patent Information
- Application Number
- CN202410904991.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-08
- Publication Date
- 2026-10-23
- Estimated Expiration
- 2044-07-08
AI Technical Summary
这就产生了大量的无用的计算消耗，使得EMTR方法的时效性有所欠缺
[0019]1.利用神经网络模型对辐射源信号的时频信息进行深度学习，挖掘辐射源与噪声信号时频域上的本质区别，进而挑选保留辐射源信号，剔除噪声信号，克服了现有的基于统计特征阈值筛选的方法无法区分噪声和弱辐射源的弊端，有助于对弱辐射源定位结果进行保留，从而更加丰富了闪电VHF成像结果图的信息。
Smart Images
Figure CN118885830B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of signal processing technology, specifically relating to a neural network pre-selection method for locating lightning VHF radiation source signals. Background Technology
[0002] Refined localization of very high frequency (VHF) lightning sources can effectively help analyze the development process of lightning channels and the spatiotemporal location of lightning sources. Compared with solving for the source location, the validity determination of the results is also a crucial step, as it determines whether the localization results for data within a certain time window are reasonable and reliable.
[0003] Recently, in pursuit of high positioning performance under noisy conditions, the electromagnetic time reversal (EMTR) method has received widespread attention and research. It can effectively capture weak radiation sources and has high positioning accuracy. Currently, two indicators are commonly used to determine the effectiveness of positioning results: coherence ratio (CR) and energy ratio (ER).
[0004] The problem lies in the fact that both of these indicators require a manually set threshold for filtering, which is roughly estimated based on the statistical distribution of the data. In reality, the ER values of many weak radiation sources are not high, comparable to or even lower than noise. To obtain a clearer lightning image, a higher ER threshold needs to be set, which inevitably leads to the filtering out of many weak radiation sources. Conversely, to obtain richer radiation source information, the ER value is set lower, which inevitably introduces noise contamination, affecting the analysis and interpretation of the channel. The reason for these problems is that both CR and ER indicators are subjectively set based on statistical data, without distinguishing between noise and radiation sources from the fundamental time-frequency characteristics of the signal. Furthermore, for a single lightning strike, only about 10% of the radiation source location results are truly valid; the rest are noise-interference results. This results in a large amount of useless computational consumption, making the EMTR method less timely. Summary of the Invention
[0005] The purpose of this invention is to provide a neural network pre-selection method for lightning VHF radiation source location signals that can quickly and accurately distinguish between radiation source and noise data, so that weak radiation source location points can be preserved.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] A neural network-preselected VHF lightning radiation source localization system includes a signal preprocessing part, a radiation source preselection part, and a localization imaging part. The signal preprocessing part uses a sliding window to extract time-domain waveform data from N antennas and performs discrete Fourier transform to obtain frequency-domain information, and obtains the imaginary and real part values and phase values of J frequency points as subsequent input data.
[0008] The radiation source pre-selection part includes a one-dimensional convolutional neural network model based on the discrete Fourier transform of the signal. The one-dimensional convolutional neural network model is used to obtain the input data of the signal preprocessing part, make predictions and output the judgment results, and retain the time-domain waveform data extracted by the sliding window that is determined to be a radiation source for subsequent positioning.
[0009] The positioning and imaging section uses the electromagnetic time reversal (EMTR) algorithm to perform two-dimensional direction of arrival positioning on the preserved time-domain waveform data, and uses a density clustering algorithm to remove noise from the initial positioning results.
[0010] The neural network pre-selection method for locating lightning VHF radiation sources using the above-described system includes the following steps:
[0011] The first step is to build the model: First, a one-dimensional convolutional neural network model based on the discrete Fourier transform of the signal is constructed for binary classification to identify radiation sources. The input of the model includes data from three channels, which are the real part, imaginary part, and phase information of J different frequency points in the frequency domain after the discrete Fourier transform of the signals received by multiple VHF array antennas. The data from each of the three channels can be processed by two one-dimensional convolutional layers. The output features can be combined by concatenation. The middle of the model includes channel and spatial attention modules to enhance the saliency of features in different channels and different frequency bands. The tail of the model is used to expand the features into a one-dimensional vector and feed it into a three-layer fully connected structure for classification prediction. Finally, the class is determined and output through the SoftMax layer. The model output is a single value, where a value of 1 represents that the input data is a radiation source and an output value of 0 represents that the input data is noise.
[0012] The second step is to train the model: Select multiple radiation source data and noise data from the VHF data of artificially triggered lightning located by the existing EMTR method, and label these data respectively. The radiation source data is labeled with label 1 and the noise data is labeled with label 0. Use these data as the training set of the above neural network model, feed the training set data into the model for training, start the training program, train for multiple iteration cycles, and after the training stops, select the model with the highest verification accuracy as the final model for use.
[0013] The third step is to preprocess the newly acquired lightning VHF data using the signal preprocessing section, then send it to the radiation source preselection section to retain the radiation source data and remove noise data. Finally, the positioning and imaging section is used to position and image the retained radiation source data.
[0014] Furthermore, the third step of localization imaging specifically involves: using EMTR to perform two-dimensional arrival direction localization on the retained radiation source data to obtain initial results, and then using a density clustering algorithm to remove noise from the initial results to achieve localization imaging.
[0015] Furthermore, the second step selected 300,000 radiation source data and 300,000 noise data.
[0016] Furthermore, the training iteration cycle in the second step is 30.
[0017] Furthermore, the signal preprocessing section performs Discrete Fourier Transform (DFT) on the time-domain waveform data of the seven antennas extracted by the sliding window. The sampling rate of the antenna time-domain waveform data is 500MHz, the data length of the sliding window is 256 sampling points, and the total data length is 512ns. The window overlap rate is 0.75. The frequency domain information is obtained using Fast Fourier Transform (FFT), yielding the imaginary and real parts and phase values of 80 frequency points in the 25-88MHz and 108-200MHz ranges, respectively, which serve as subsequent inputs.
[0018] Compared with the prior art, the present invention has the following effects:
[0019] 1. By using a neural network model to perform deep learning on the time-frequency information of radiation source signals, the essential differences between radiation source signals and noise signals in the time-frequency domain are explored. This allows for the selection and retention of radiation source signals while eliminating noise signals. This overcomes the shortcomings of existing methods based on statistical feature thresholds, which cannot distinguish between noise and weak radiation sources. It helps to preserve the location results of weak radiation sources, thereby enriching the information in the lightning VHF imaging results.
[0020] 2. A large number of noisy sliding windows were pre-screened, reducing the computational load of the EMTR localization process and improving the computational speed of lightning localization imaging. Attached Figure Description
[0021] Figure 1 This is a general operational block diagram of the present invention;
[0022] Figure 2 yes Figure 1 A schematic diagram of the neural network structure of a CNN prediction model;
[0023] Figure 3 yes Figure 1A schematic diagram of the medium-density clustering module;
[0024] Figure 4 This is the imaging image obtained by the method of the present invention for positioning;
[0025] Figure 5 This is an image obtained from the existing EMTR localization algorithm;
[0026] Figure 6 This is an image of the radiation source location point obtained by the method of the present invention compared to existing methods.
[0027] Figure 1 The algorithm consists of 1. Sliding window, 2. Neural network input, 3. CNN prediction model, 4. Electromagnetic time reversal positioning algorithm, 5. Density clustering algorithm, 6. Signal preprocessing, 7. Radiation source preselection, and 8. Positioning and imaging. Detailed Implementation
[0028] The implementation of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.
[0029] Combination Figure 1 A neural network-preselected VHF lightning radiation source localization system includes a signal preprocessing part 6, a radiation source preselection part 7, and a localization imaging part 8. The signal preprocessing part 6 uses a sliding window 1 to extract the time-domain waveform data of 7 antennas and performs discrete Fourier transform to obtain frequency domain information, and obtains the imaginary and real part values and phase values of 80 frequency points as subsequent input data.
[0030] The radiation source pre-selection part 7 includes a one-dimensional convolutional neural network model 3 based on the discrete Fourier transform (DFT) of the signal. The one-dimensional convolutional neural network model 3 is used to obtain the input data of the signal preprocessing part 6, make predictions and output the judgment results, and retain the time-domain waveform data extracted by the sliding window 1 that is determined to be a radiation source for subsequent positioning.
[0031] The positioning and imaging section 8 uses an electromagnetic time reversal algorithm to perform two-dimensional arrival direction positioning on the retained time-domain waveform data, and uses a density clustering algorithm to remove noise from the initial positioning results.
[0032] The neural network pre-selection method for locating lightning VHF radiation sources using the above-described system includes the following steps:
[0033] The first step is to build a model: combining Figure 2First, a one-dimensional convolutional neural network model 3 based on the discrete Fourier transform of the signal is constructed for binary classification to identify radiation sources. The input of the model includes data from three channels, namely the real part, imaginary part, and phase information of 80 different frequency points in the frequency domain after the discrete Fourier transform of the signals received by the 7 VHF array antennas. The data from each of the three channels can be processed by two one-dimensional convolutional layers (Conv), and the output features can be combined by concatenation. The middle of the model includes channel and spatial attention modules to enhance the saliency of features in different channels and different frequency bands. The tail of the model is used to flatten the features into a one-dimensional vector and feed it into a three-layer fully connected structure (Dense) for classification prediction. Finally, the class is determined and output through the SoftMax layer. The model output is a single value, where a value of 1 represents that the input data is a radiation source and an output value of 0 represents that the input data is noise.
[0034] The second step is to train the model: Select multiple radiation source data and noise data from the VHF data of artificially triggered lightning located by the existing EMTR method, and label these data respectively. The radiation source data is labeled with label 1 and the noise data is labeled with label 0. Use these data as the training set of the above neural network model, feed the training set data into the model for training, start the training program, train for multiple iteration cycles, and after the training stops, select the model with the highest verification accuracy as the final model for use.
[0035] Step 3: The newly acquired lightning VHF data is preprocessed by the signal preprocessing section 6 and then sent to the radiation source preselection section 7 to retain the radiation source data and remove noise data. The retained radiation source data is then located and imaged by the positioning and imaging section 8.
[0036] Furthermore, in the third step, the localization imaging specifically involves: the localization imaging part 8 using the EMTR algorithm to perform two-dimensional arrival direction localization on the retained radiation source data to obtain initial results, and then using a density clustering algorithm to remove noise from the initial results obtained from the localization, thereby achieving localization imaging.
[0037] Furthermore, the second step selected 300,000 radiation source data and 300,000 noise data.
[0038] Furthermore, the training iteration cycle in the second step is 30.
[0039] Furthermore, the signal preprocessing section 6 performs Discrete Fourier Transform on the time-domain waveform data of the 7 antennas extracted by the sliding window 1. The sampling rate of the antenna time-domain waveform data is 500MHz, the data length of the sliding window 1 is 256 sampling points, and the total data length is 512ns. The window overlap rate is 0.75. The frequency domain information is obtained by using the Fast Fourier Transform method, and the imaginary and real part values and phase values of 80 frequency points in the frequency domain signals of 25-88MHz and 108-200MHz are obtained as subsequent inputs.
[0040] Furthermore, the localization imaging section employs EMTR algorithm 4 to perform two-dimensional arrival direction localization on the retained sliding window data. Density clustering algorithm 5 is used to remove noise from the initial localization results; the principle diagram of density clustering is shown below. Figure 3 For spatially distributed points, this algorithm divides the points into core points, boundary points, and noise points by setting a neighborhood radius ε and a minimum number of neighborhoods m. In this system, the neighborhood radius ε is set to 2° and the minimum number of neighborhoods m = 4. Noise points are removed, and core points and boundary points are retained as the final localization results.
[0041] Figure 4-6 This is a two-dimensional (azimuth and elevation) localization image of the VHF lightning radiation source. The entire VHF lightning data was divided into 3,898,400 sliding window data points. Figure 4 The image obtained by the method of the present invention retains a total of 443,024 radiation source location points; Figure 5 The existing EMTR localization algorithm yielded 305,390 localization points in the resulting image. The method of this invention locates 138,598 more radiation sources than existing methods, representing an improvement of 45.38%. Furthermore, compared to existing methods, 946 radiation source localization points were missed, representing a miss rate of 0.31%. The overall improvement in radiation source localization is 45.07%. Figure 6 The image shows the additional radiation source location points obtained by the method of this invention compared to existing methods. By comparison, it can be seen that the image obtained by this method has richer details.
[0042] The method of this invention takes a total of 64 hours to locate, while the existing EMTR method takes a total of 456 hours, which is 7.1 times faster.
[0043] This invention can pre-screen time-domain signals and identify and retain weak radiation source location results, thereby improving the quality and efficiency of lightning location imaging.
[0044] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A neural network-based VHF lightning radiation source localization system, characterized in that, It includes a signal preprocessing part (6), a radiation source preselection part (7), and a positioning and imaging part (8). The signal preprocessing part (6) uses a sliding window (1) to extract the time-domain waveform data of N antennas and performs discrete Fourier transform to obtain frequency domain information, and obtains the imaginary and real part values and phase values of J frequency points respectively, as subsequent input data. The radiation source pre-selection part (7) includes a one-dimensional convolutional neural network model (3) based on the discrete Fourier transform of the signal. The one-dimensional convolutional neural network model (3) is used to obtain the input data of the signal preprocessing part (6), make predictions and output the judgment results, and retain the time-domain waveform data extracted by the sliding window (1) that is determined to be a radiation source for subsequent positioning. The construction method of the one-dimensional convolutional neural network model (3) is as follows: the one-dimensional convolutional neural network model (3) based on the discrete Fourier transform of the signal is constructed for binary classification to identify radiation sources. The input of the model includes data from three channels, which are the real part, imaginary part and phase information of J different frequency points in the frequency domain after the signal received by multiple VHF array antennas is discrete Fourier transformed. The data from the three channels can each pass through two one-dimensional convolutional layers. The output features can be combined together through splicing operations. The middle of the model includes channels and spatial attention modules to enhance the saliency of features in different channels and different frequency bands. The tail of the model is used to expand the features into a one-dimensional vector and send it into a three-layer fully connected structure for classification prediction. Finally, the category is determined and output through the SoftMax layer. The model output is a single value. The value 1 represents that the input data is a radiation source, and the output value 0 represents that the input data is noise. The positioning and imaging section (8) uses the electromagnetic time reversal algorithm to perform two-dimensional arrival direction positioning on the retained time-domain waveform data, and uses the density clustering algorithm to remove noise from the initial positioning results.
2. A neural network pre-selection method for locating lightning VHF radiation source signals using the system described in claim 1, characterized in that, Includes the following steps: The first step is to train the model: select multiple radiation source data and noise data from the VHF data of artificial lightning located by the existing EMTR method, and label these data respectively. The radiation source data is labeled with label 1 and the noise data is labeled with label 0. These data are used as the training set of the above one-dimensional convolutional neural network model (3). The training set data is sent into the model for training. Start the training program and train for multiple iteration cycles. After the training stops, select the model with the highest verification accuracy as the final model to be used. Step 2: The newly acquired lightning VHF data is preprocessed by the signal preprocessing part (6) and then sent to the radiation source preselection part (7) to retain the radiation source data, remove noise data, and then the retained radiation source data is positioned and imaged by the positioning imaging part (8).
3. The method according to claim 2, characterized in that, The second step of positioning and imaging specifically involves: the positioning and imaging part (8) uses the frequency domain EMTR method to perform two-dimensional arrival direction positioning on the retained radiation source data to obtain the initial result, and uses the density clustering algorithm to remove noise from the initial result obtained by positioning to achieve positioning and imaging.
4. The method according to claim 2 or 3, characterized in that, The first step selected 300,000 radiation source data and 300,000 noise data.
5. The method according to claim 2 or 3, characterized in that, The first step involves 30 training iterations.
6. The method according to claim 2 or 3, characterized in that, The signal preprocessing part (6) performs discrete Fourier transform on the time-domain waveform data of the 7 antennas extracted by the sliding window (1). The sampling rate of the antenna time-domain waveform data is 500MHz. The data length of the sliding window (1) is 256 sampling points, with a total data length of 512 ns. The window overlap rate is 0.
75. The frequency domain information is obtained by using the fast Fourier transform method. The imaginary and real part values and phase values of 80 frequency points in the frequency domain signal from 25-88 MHz and 108-200 MHz are obtained respectively, which are used as subsequent inputs.